Australian control of AI inference

Australian control of AI inference

Personal submission to the Department of the Prime Minister and Cabinet

Alex Carpenter | 30/09/2026

Consultation: Getting it right: Building AI infrastructure that works for Australia

Purpose of this submission

Thank you for the opportunity to contribute. I am a PhD researcher studying how Australian small and medium enterprise leadership teams frame and respond to AI. I make this submission in a personal capacity.

My central recommendation is that every Australian business should have a practical option to use capable AI models served within Australia under Australian control. Businesses should remain free to use overseas services. Sovereignty should give us a credible choice and continuity of service, supported by the skills to use that choice well.

I support developing Australia's AI infrastructure and working with international developers. I also believe government needs to invest substantially in the capability it wants Australia to control. The standards should connect those ambitions to an outcome that businesses can actually use.

Conditions for AI training

Sovereignty requires control of inference

The consultation distinguishes AI training from inference, which is the use of a trained model to produce an output. It places considerable emphasis on attracting frontier training to Australia as a source of sovereign capability and economic benefit (Department of the Prime Minister and Cabinet [PM&C], 2026).

For sovereignty, my priority is inference. What matters to me is whether Australian organisations can run and use capable models here under Australian control, wherever those models were trained. Domestic training may bring research, employment and other economic benefits. Those benefits deserve consideration, but training a model here does not by itself establish Australian control over its subsequent use.

The same applies to the presence of data centres. An Australian building and local data storage are insufficient if an overseas organisation retains the ability to withdraw the model or the services needed to operate it. The relevant test is whether an Australian operator can legally and practically continue running an agreed model if its overseas partner withdraws access.

The proposed conditions for frontier training should be complemented by an explicit approach to sovereign inference. Different arrangements should be able to meet that outcome, rather than assuming a particular ownership structure or technical design. Continuity needs to be credible in practice. A contractual promise should not be treated as proof that an overseas legal direction could never interrupt service.

Overseas decisions can interrupt access

Recent events show why this matters. On 12/06/2026, Anthropic reported that a US export-control directive restricting foreign-national access required it to suspend Fable 5 and Mythos 5 for all customers. Other Anthropic models remained available. The company subsequently announced the lifting of those export controls and the restoration of global Fable 5 access from 01/07/2026 (Anthropic, 2026a, 2026b). This was a temporary interruption affecting particular models, rather than a general shutdown of AI services. It nevertheless demonstrates that access can change abruptly because of decisions taken overseas.

A separate decision on 25/09/2026 provides another example of government intervention in AI supply. By a two-to-one majority, the US Court of Appeals for the District of Columbia Circuit upheld the exclusion of Claude from the Department of War's supply chain under the Federal Acquisition Supply Chain Security Act. That ruling concerned defence procurement. It did not authorise a general ban on public access to AI (Anthropic PBC v. United States Department of War, 2026).

Australia's own Digital Transformation Agency already recognises the broader issue. Its coordinated contracting principles consider whether foreign laws or directions could override Australian contractual protections, who holds operational and decision-making authority, and whether services could be removed. They also address continuity, portability and practical exit arrangements. These are Commonwealth procurement principles, but the underlying questions are relevant to Australian businesses too (Digital Transformation Agency [DTA], n.d.).

My concern is reducing Australia's exposure to decisions we cannot control. I would like the standards to make a dependable Australian inference option part of the response, without suggesting that any contract or bilateral agreement can simply override foreign law.

Preserve international partnerships and model choice

Australia should remain open to partnerships with US and Chinese developers, alongside Australian providers. That includes working with companies such as OpenAI and Anthropic. We should be able to use frontier commercial services and open-weight models, with Australian sovereignty and security requirements applied to the particular arrangement.

Open weights should not be treated as automatic sovereignty. The rights and practical ability to operate a model still matter. Equally, an overseas developer should not be ruled out simply because it is overseas. Australia should assess whether the proposed arrangement delivers the control and security required, and refuse particular models or deals where the risks are unacceptable.

I am proposing an available Australian option. Businesses should not be compelled to use it or prevented from choosing an overseas service that better suits their work. Some may reasonably pay a premium for Australian control and continuity; others may make a different choice.

Invest in local capability skills research and innovation

Practical implementation, training and good working practices are essential to sovereignty. Access to a model does not establish that an organisation can use it appropriately, efficiently or effectively. People need to understand how to build useful workflows, judge outputs and manage the responsibilities that come with using AI in their work.

The consultation already recognises that compute access needs accompanying models, storage, networking and technical support (PM&C, 2026). Developer contributions should therefore be assessed by whether Australian organisations can put that capability to work. Nominal access to computing resources is a weak outcome if businesses lack the support needed to use them.

International evidence reinforces this distinction. The OECD's 2025 report on SMEs, based on a survey in seven countries that did not include Australia, found that a third or fewer of businesses using generative AI had taken measures such as staff training, internal guidelines or investigating legal issues. Respondents welcomed training support (OECD, 2025). This is not an estimate of Australian practice, but it supports taking implementation capability seriously alongside access.

I would like Australian businesses to be able to develop everyday AI workflows using services under sovereign control, while retaining access to foreign models. Contributions from developers and infrastructure providers should include practical assistance and skills development that make this possible. The standard should be whether the support helps organisations use the capability well, rather than whether a provider can report a nominal allocation of compute.

Requirements for large data centres

Public investment and consumer costs

If government wants meaningful Australian control, it needs to put substantial, credible funding behind that ambition. A shared public and private investment model makes sense. Australia cannot expect other organisations to pay for everything that Australia intends to control.

This includes the electricity and water infrastructure needed to support the capability. I accept both direct government investment and some shared infrastructure costs through household and business electricity and water bills. Those contributions should be explicit and justified by the capability and infrastructure they help build. The public should be able to understand what it is paying for and why.

This differs from the consultation's objective of imposing no net costs on consumers and communities (PM&C, 2026). I support careful scrutiny of costs and a fair contribution from developers. I do not support making the absence of any public or shared contribution an absolute condition where Australia is seeking a lasting sovereign benefit.

I draw a clear distinction between investing in sovereign capability and subsidising AI consumption. I do not support general subsidies for AI usage. Businesses and other users should pay for the AI services they consume, including a possible premium for an Australian-controlled service. Public investment should help establish the capability and control that Australia needs; it should not become a general discount on users' AI bills.

Make a positive contribution to the energy transition

I strongly support new renewable generation to match data-centre electricity demand. AI infrastructure investment presents an opportunity to accelerate Australia's renewable grid development. Government, utilities and developers should use that opportunity to build the supporting infrastructure and move viable projects forward.

The energy requirements should give proponents a practical route to expanding AI capability and renewable supply together.

Use proportionate water requirements

Water use needs to be assessed in context. Cooling design, the water source and local supply conditions matter. The Water Services Association of Australia's review describes different cooling approaches and the need to consider their local implications (Water Services Association of Australia [WSAA], 2025).

Recycled supply offers a concrete pathway. In June 2026, Greater Western Water announced that its Melton plant would supply recycled cooling water to a planned AWS data centre in western Melbourne from the beginning of operations. This is a planned arrangement, not evidence of savings already achieved (Greater Western Water, 2026).

Water requirements should be proportionate to the proposed facility's actual demand, cooling design and local supply. Low-water designs and suitable recycled-water supplies should enable timely development. Where a genuine local constraint exists, governments, utilities and developers should work through practical solutions, including shared infrastructure or a more suitable location. Water concerns should not become a pretext for unnecessary delay. The economic cost of delaying Australian AI capability also needs to be taken seriously.

Treat delay as a cost

Delay leaves Australia dependent on decisions made elsewhere and postpones economic opportunities. In my view, the economic opportunities forgone are the larger concern. That does not make every delay unjustified, but it means the cost of waiting belongs in the assessment alongside the risks of proceeding.

I favour practical compromise, common sense and flexibility about location. Where moving a project would resolve a real constraint, that should be pursued. Where a workable solution exists, governments should help deliver it promptly. The standards should make responsible development achievable.

Recommendations

I recommend that the Australian Government:

• Make access to inference served within Australia under Australian control an explicit sovereign outcome, available as a practical option to every Australian business.
• Assess sovereignty through the legal and practical ability to continue operating an agreed model, while leaving the means of delivering that outcome open.
• Preserve overseas-provider choice and pursue international partnerships, subject to Australian security requirements and the ability to refuse unacceptable arrangements.
• Treat implementation support, training and sound working practices as essential components of Australian AI capability.
• Commit substantial public investment alongside private investment, including a transparent share of supporting electricity and water infrastructure costs, while leaving users responsible for the AI services they consume.
• Support new renewable generation to match demand, apply proportionate water requirements and take the economic cost of unnecessary delay seriously.

References

Anthropic. (2026a, 30/06/2026, updated 01/07/2026). Redeploying Fable 5.

Anthropic. (2026b, 12/06/2026). Statement on the US government directive to suspend access to Fable 5 and Mythos 5.

Anthropic PBC v. United States Department of War, No. 26-1049 (D.C. Cir. 25/09/2026). Court opinion.

Department of the Prime Minister and Cabinet. (2026). Getting it right: Building AI infrastructure that works for Australia: Consultation paper - September 2026.

Digital Transformation Agency. (n.d.). Coordinated contracting arrangements (CCA) principles. Retrieved 30/09/2026.

Greater Western Water. (2026, 18/06/2026). Melton plant leads the way with first recycled water connection for a Victorian data centre.

OECD. (2025). Generative AI and the SME workforce: New survey evidence. OECD Publishing.

Water Services Association of Australia. (2025). Data centres and water in Australia.